برمجة Perceptron من الصفر

A perceptron is one neuron plus a learning rule. It is the original trainable building block of neural networks, and you can code it from scratch. 🛠️

First create the weights and a bias. Starting them at zero is fine for a perceptron; learning will move them where they need to be.

The predict step computes the weighted sum, then applies the step function to return 0 or 1.

Compare the prediction to the true label. The error is simply target minus prediction: 0 when right, plus or minus one when wrong.

The update rule nudges each weight by the error times the input. Wrong guesses push the weights toward the correct answer.

The bias learns the same way, but its input is always 1. So you simply add the learning rate times the error.

The learning rate scales how big each update is. Small values learn slowly but steadily; large ones can overshoot the answer.

Looping once over every training example is called an epoch. You usually repeat for several epochs until mistakes stop.

Each step of training predicts, measures error, and updates the weights and bias. Repeat across the dataset to learn.

Feed it the AND truth table and the perceptron converges fast. Both inputs must be 1 for it to output 1.

If the data is linearly separable, the perceptron is guaranteed to find a separating line. That promise is the convergence theorem.

A perceptron predicts with a step, measures error, and updates weights by lr times error times input. Repeat over epochs and it learns separable data. ✅